How would you measure the success of Composer 2, Cursor's in-house coding model?

Introduction

Composer 2 is Cursor's own coding model, launched on 19 March 2026, two weeks before Cursor 3. It is a mixture-of-experts model trained with reinforcement learning inside Cursor's agent environment, with tools such as file editing, terminal commands, grep and semantic codebase search. Cursor positioned it as close to frontier quality on coding tasks at a fraction of the price, reported at about $0.50 per million input tokens and $2.50 per million output tokens. It also drew criticism when users discovered it was built on Moonshot AI's open-weight Kimi base model before Cursor said so.

Measuring the success of a model is different from measuring a feature. A model can top benchmarks and still fail in the product, or be slightly weaker but far cheaper and faster, which might be exactly what Cursor needs. I'll define success across user value, model quality, unit economics and trust.

Step 1

Product Context (3 minutes)

Cursor sells access to many models: OpenAI, Anthropic and Google frontier models, plus its own Composer family. Paid plans include a usage pool billed at model API rates, and Auto mode chooses a model for the user. Every request to a third-party frontier model costs Cursor money paid to a competitor that also sells its own coding tools (Claude Code, Codex, Gemini CLI).

Composer 2 exists for four reasons:

• Margin: serving its own model is cheaper than paying frontier API rates, which matters as ARR passed a reported ~$3B in 2026.
• Speed: a model tuned for Cursor's agent loop can be faster per task.
• Independence: less reliance on suppliers who are also rivals.
• Product fit: the model is trained on Cursor's real tool environment, so it can use Cursor's tools better than a general model.

Key stakeholders: developers (want good, fast results), Cursor (margin, independence), enterprise buyers (care about security, data use and model provenance), and model partners.

Where it shows up: Auto mode, the model picker in Agent, the Agents Window and cloud agents.

Lifecycle stage: growth. The model is launched; the question is whether it becomes the default for most agent work.